I was chatting recently with a colleague who had just returned from a financial services conference. He described a demo where an AI agent – not a human, an ‘autonomous’ software agent – had negotiated a series of foreign exchange (FX) trades, routed payments across three correspondent banks, flagged a potential AML concern in the middle of the sequence, paused the workflow to log the flag, and then resumed and settled the whole transaction chain. All in under four seconds. The weird thing about the demo was that nobody touched a keyboard. His reaction was equal parts impressed and unsettled. Mine was: what happens when that agent’s infrastructure goes down somewhere in the middle of all that?
That question is what this article is about.
What Do We Mean by Agentic AI?
Most of the AI we have discussed in previous editions of Trends and Wins has been what I would call ‘assistive’ AI. A model analyzes data, makes a recommendation or prediction, and a human decides what to do with it. The robo-advisors I mentioned in the AI in Financial Services article (issue 17) are a good example. They suggest. You approve.
Agentic AI is a different animal. An AI agent doesn’t just recommend – it acts. It is given a goal and a set of tools (APIs, databases, messaging systems, other AI models) and it autonomously plans and executes a sequence of steps to achieve that goal. It can call other agents, spawn sub-tasks, loop back when something goes wrong, and make real-time decisions along the path. Think of it less like a smart assistant and more like a junior employee you have handed a workflow to and said, ‘go handle it.’
The industry has moved here faster than most expected. OpenAI, Anthropic, Google, and virtually every major technology vendor now offers agentic frameworks. Enterprise adoption is accelerating, particularly in financial services, healthcare, and logistics – the
very industries that have historically depended on Nonstop systems. This is not a future trend. Agentic AI is being deployed in production environments today.
The Infrastructure Question Nobody Is Asking
There is an enormous amount of discussion in the industry right now about which AI models are most capable, which agent frameworks are best, and how to govern AI decision-making. All important questions. But there is a question being asked far less
often: what happens to a running agentic workflow when the underlying infrastructure fails?
With traditional applications, failure has a relatively clean story. The application stops. Humans notice. The team restores the system. Transactions that were in-flight may need to be investigated and reconciled. This is painful but understood. Nonstop was built precisely to eliminate this failure mode for the most critical workloads, and our track record speaks for itself.
With agentic AI, the failure story is considerably more complicated. An agent might be mid-way through a multi-step process: it has already called three APIs, written to two databases, sent a message to a downstream system, and is now evaluating the response from a credit scoring model before deciding its next action. If the infrastructure fails at that moment, what exactly is the state of the world? The agent has already taken actions. Some of those actions may be irreversible. A trade has been placed. A payment has been initiated. A regulatory flag has been logged. Or perhaps not – because the write failed silently and the agent never knew.
This is the agentic AI infrastructure problem, and it is not a small one.
Why This Is a Nonstop Story
Regular readers of this column know I occasionally get accused of finding a way to bring everything back to Nonstop. Guilty as charged. But in this case, I think the connection is not a stretch – it is essentially the point.
The Nonstop system was designed, from the ground up, to run processes that cannot be interrupted. The architecture – process pairs, message-based communication, fault-tolerant storage, single-system image across multiple processors – was built around the premise that certain workloads are simply too important to fail. Payment authorization. Clearing and settlement. The systems where an outage is not an inconvenience but a
crisis.
Agentic AI in financial services is heading directly into that territory. When an AI agent is autonomously executing payment sequences, managing risk positions, or processing loan approvals in real time, it inherits all of the availability requirements of those
underlying transactions. The fact that a software agent is doing the work rather than a human operator does not make the work any less critical. If anything, the speed and autonomy of the agent raises the stakes, because errors can propagate faster than any
human could catch them.
What does a mission-critical agentic AI platform require? Let me suggest a few fundamentals:
Continuous availability – 24x7x365, no planned or unplanned downtime. An agent workflow that is interrupted mid-execution is potentially worse than one that never started, because of the partial state problem described above.
Transactional integrity – every action the agent takes must be auditable, recoverable, and consistent. If an agent writes to a database and the system fails, there can be no ambiguity about whether the write succeeded. This is ACID-level commitment for AI
workflows.
Fault-tolerant execution – the agent runtime itself needs to survive hardware and software failures without losing context. In Nonstop terms, this is the process-pair model applied to AI orchestration.
Scalability without disruption – agentic workloads will be spiky. A market event, a regulatory deadline, a fraud spike – these will drive sudden surges in agent activity. The infrastructure needs to be scaled immediately and transparently, without taking anything
offline. We have seen this story before. I think back to the Covid unemployment processing story from issue 9, where Nonstop scaled out over a weekend to handle a 20x surge in volume without missing a beat. That capability becomes even more
important when the workloads are autonomous and moving fast.
Auditability – regulators are going to want a complete record of what an AI agent decided, why it decided it, what actions it took, and in what sequence. This is not optional in financial services. The audit trail must be immutable and complete, which is
another way of saying it must live on infrastructure that cannot be corrupted or interrupted.
The Accountability Dimension
I want to say a word about accountability, because I think it is underappreciated in the current AI conversation. When a human operator makes an error in a financial workflow, there is a person accountable for that error. When an AI agent makes an error,
accountability is more diffuse – was it the model? The training data? The agent framework? The business logic that defined the agent’s goals? The infrastructure that failed to record what happened?
Infrastructure that keeps a perfect, tamper-proof record of every agent action is not just an operational nicety – it is the foundation of accountability in an agentic world. Without it, when something goes wrong (and something will go wrong), the organization has no undisputable truth to stand on.
In my article on Machine Unlearning (issue 29), I wrote about the importance of being able to trace and if necessary, reverse what an AI model has learned. The same principle applies here at the operational level. We need to be able to trace and if necessary, reverse what an AI agent has done. That requires an infrastructure that never loses a record.
Where We Go From Here
I think the agentic AI wave is going to force a reckoning in enterprise infrastructure. For the past decade, the conversation in most data centers has been about moving workloads to the public cloud, embracing microservices, and accepting that occasional failures are normal and recoverable. That philosophy works fine when humans are in the decision loop and the consequences of a brief outage are modest.
But as AI agents take on more autonomous decision-making in genuinely mission-critical workflows, the tolerance for failure will collapse back toward zero. Not because organizations will suddenly fall in love with on-premise infrastructure, but because the business and regulatory consequences of an agent failure in a live financial process are
simply too severe.
That is where the decades of work that has gone into Nonstop start to look very relevant again. The problems that Tandem set out to solve in the 1970s – how do you build a computer system that keeps running no matter what? – are the same problems that agentic AI in critical industries is going to impose on infrastructure architects in the coming years. The answers, refined over fifty years and still being refined, are not going to become less valuable. They are going to become more valuable.
My friend from the conference eventually settled his unease about the FX trading demo by noting that the AI was running on a highly redundant cloud environment with failover. I did not have the heart to tell him that ‘highly redundant’ and ‘Nonstop’ are not the same thing. But that conversation is coming, and I think it is going to be a good one for us.
As always, I welcome your thoughts and reactions. Email me at
justin.simonds@hpe.com.
Be the first to comment